bm-mistral-7b-transcription-correction

Bambara ASR post-correction: djelia/bm-mistral-7b-v1 specialised on rewriting raw speech-recognition output into correct Bambara. It repairs split words, wrongly merged words, mis-transliterated French loanwords, and dropped words or characters, while preserving meaning and orthography.

MistralForCausalLM, bfloat16 — 32 layers, hidden size 4096, 32 heads with 8 KV heads (GQA), 32,768-token vocabulary and context.

Usage

Load the merged weights with AutoModelForCausalLM; the repo also ships the LoRA adapter. There is no chat template — the instruction goes under ### ɲɛfɔli:, the raw transcript under ### Donnafɛnw:, and generation starts after ### Jaabi:.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "djelia/bm-mistral-7b-transcription-correction"
tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.padding_side = "left"  # repo ships "right"; left is required for batched generation
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

INSTRUCTION = """I ye Bambara sɛbɛnni kɔrɔsibaga ye min bɛ ASR Bambara sɛbɛnni ɲɛnabɔ. I ka baara ye ka Bambara sɛbɛnni fili minnu bɛ ASR la, olu yɛlɛma ka kɛ Bambara sɛbɛnni ɲuman ye, ka kɔrɔ bɛɛ to a cogo la. I ka kan ka fili suguyaw ninnu ɲɛnabɔ:

Daɲɛw Tilili: Tuma dɔw la, ASR bɛ daɲɛ kelen tila ka kɛ daɲɛ fitini caman ye. I ka kan ka olu fara ɲɔgɔn kan ka kɛ daɲɛ ɲuman kelen ye.
Daɲɛw Farali: Tuma dɔw la, daɲɛ fla bɛ fara ɲɔgɔn kan ka kɛ kelen ye. I ka kan ka olu tila ka Bambara sɛbɛnni cogo ɲuman bato.
Tubabukan Yɛlɛmali Fili: Tubabukan kumaw bɛ se ka yɛlɛma Bambara la ni kanfɔ suguya wɛrɛ ye min tɛ a ɲuman ye (misali la, "cette fois-ci" bɛ se ka kɛ "se ti fassi si" ye walima "à travers" bɛ kɛ "a taara were" ye). I ka kan ka olu sɛbɛn cogo ɲuman na.
Daɲɛw walima Sɛbɛndenw Tununi: Tuma dɔw la, daɲɛw walima sɛbɛnden kelenkelenna dɔw bɛ bɔ sɛbɛnni na.

Ni i bɛ Bambara kumakan dɔ ɲɛnabɔ, i ka kan ka fili suguyaw ninnu bɛɛ ɲɛnabɔ ka sɔrɔ ka kɛ Bambara sɛbɛnni ɲuman ye, nka ka kanfɔcogo bato walasa ka bɛn ni fɔcogo ye.
I ka labaaraw ka kan ka kɛ Bambara sɛbɛnni ɲɛnabɔlen ye, ni daɲɛw danw, tomi, ani daɲɛ sugandilen ɲumanw ye minnu bɛ bɛn ni kanfɔcogo ye. I ka jija ka kɔrɔ fɔlen to a cogo la ka sɔrɔ ka a ɲɛfɔ ka ɲɛ ani ka a kɛ sɛbɛnni ɲuman ye."""

ALPACA_PROMPT = """Nin ye baara dɔ ɲɛfɔli ye, min bɛ donnafɛnw ni sigidaw fara ɲɔgɔn kan. I ka kan ka jaabi sɛbɛn min bɛ ɲinini dafa ka ɲɛ.

### ɲɛfɔli:
{}

### Donnafɛnw:
{}

### Jaabi:
"""

prompt = ALPACA_PROMPT.format(INSTRUCTION, "<raw Bambara ASR output>")
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=300, pad_token_id=tokenizer.pad_token_id)
print(tokenizer.decode(output[0], skip_special_tokens=True).split("Jaabi:")[-1].strip())

Greedy decoding is the sensible default — this is constrained rewriting, not open generation.